dhavli

Tuning llama.cpp's Vulkan backend for AMD Vega 56/64 (gfx900/GCN5): 1-bit and MoE inference kernels, multi-GPU layer-split findings, and the dead ends